Bug 217154

Summary: Absolutely positioned and negative z-index div with canvas child gets drawn with wrong stacking order
Product: WebKit Reporter: Yong Su <jean.timex>
Component: CompositingAssignee: Rob Buis <rbuis>
Status: RESOLVED FIXED    
Severity: Major CC: bfulgham, changseok, darin, esprehn+autocc, ews-watchlist, fred.wang, glenn, kondapallykalyan, pdr, rbuis, simon.fraser, webkit-bug-importer, zalan
Priority: P2 Keywords: InRadar
Version: Safari 14   
Hardware: All   
OS: All   
See Also: https://bugs.webkit.org/show_bug.cgi?id=238589
Attachments:
Description Flags
Screenshot of incorrect stacking order
none
Patch
none
Patch
none
Patch
none
Patch ews-feeder: commit-queue-

Description Yong Su 2020-09-30 16:42:53 PDT 
Created attachment 410171 [details]
Screenshot of incorrect stacking order

Steps to reproduce:
1. Navigate to http://jsfiddle.net/jeantimex/Lsvyz0nj/
2. Notice that there are two divs (blue and red)

What to expect:
The red div should stack above the blue div.

Actual result:
The blue div stacks above the red div.

Comments:
Both blue and red divs are positioned absolutely, the blue div has a canvas child. Both divs have the same negative z-index.

This bug does not happen on Chrome, Firefox, IE, Edge. It affects all Safari browsers on all platforms.

Workarounds:
1. Change both z-index to positive, e.g., 1 (however, in some business cases, it's expected both divs have the same negative z-index).
2. Change the red view z-index to 0. (however, in some business cases, it's expected both divs have the same negative z-index).
3. Set transform: translateZ(0) or translate3d(0, 0, 0) to the red div. (I guess this will force a paint for the red div in a different layer)
Comment 1 Radar WebKit Bug Importer 2020-09-30 19:08:51 PDT 
<rdar://problem/69818585>
Comment 2 Rob Buis 2021-08-30 23:34:18 PDT 
Created attachment 436847 [details]
Patch
Comment 3 Rob Buis 2021-08-31 09:08:11 PDT 
Created attachment 436891 [details]
Patch
Comment 4 Simon Fraser (smfr) 2021-08-31 13:36:53 PDT 
Comment on attachment 436891 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=436891&action=review

> LayoutTests/ChangeLog:11
> +        * compositing/composited-canvas-overlaps-absolute-div-expected.html: Added.
> +        * compositing/composited-canvas-overlaps-absolute-div.html: Added.

I think the test name should have "multiple-negative-z" in it since that's key to the bug. And the fact that it's canvas isn't relevant (in fact, it would be better to make a test that doesn't use canvas).
Comment 5 Darin Adler 2021-08-31 13:38:54 PDT 
Comment on attachment 436891 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=436891&action=review

> Source/WebCore/rendering/RenderLayerCompositor.cpp:1101
> +    Vector<RenderLayer*> layerWillCompositeNeeded;

If this typically is a small set of layers, we should optimize by adding inline capacity:

    Vector<RenderLayer*, 2> ...

We just need a small number that is almost always higher than the number of layers.

> Source/WebCore/rendering/RenderLayerCompositor.cpp:1111
> +        if (childLayer) {

This isn’t needed. We would be adding null layer pointers to the vector.
Comment 6 Rob Buis 2021-09-01 01:18:57 PDT 
Created attachment 437002 [details]
Patch
Comment 7 Rob Buis 2021-09-01 06:28:13 PDT 
Comment on attachment 436891 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=436891&action=review

>> Source/WebCore/rendering/RenderLayerCompositor.cpp:1101
>> +    Vector<RenderLayer*> layerWillCompositeNeeded;
> 
> If this typically is a small set of layers, we should optimize by adding inline capacity:
> 
>     Vector<RenderLayer*, 2> ...
> 
> We just need a small number that is almost always higher than the number of layers.

Fair enough, however I now use a simple counter, since we do not need to actually remember the RenderLayers, just how many.

>> Source/WebCore/rendering/RenderLayerCompositor.cpp:1111
>> +        if (childLayer) {
> 
> This isn’t needed. We would be adding null layer pointers to the vector.

See above.

>> LayoutTests/ChangeLog:11
>> +        * compositing/composited-canvas-overlaps-absolute-div.html: Added.
> 
> I think the test name should have "multiple-negative-z" in it since that's key to the bug. And the fact that it's canvas isn't relevant (in fact, it would be better to make a test that doesn't use canvas).

Done.
Comment 8 Darin Adler 2021-09-01 10:02:33 PDT 
Comment on attachment 437002 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=437002&action=review

> Source/WebCore/rendering/RenderLayerCompositor.cpp:1101
> +    unsigned layerWillCompositeNeeded = 0;

Style thing: Typically it’s better to name a count with a noun like "number" or "count". A phrase like this can sound like it might be a set, not an integer. Maybe "numberOfLayerWillCompositeCallsNeeded", but that’s a bit long. Maybe "newlyCompositedChildLayerCount"?

Out of my depth, I guess, but I’ve got to admit, I am puzzled about why we need to call layerWillComposite over and over again.
Comment 9 Rob Buis 2021-09-02 00:24:27 PDT 
Created attachment 437119 [details]
Patch
Comment 10 EWS 2021-09-02 02:12:08 PDT 
Committed r281913 (?): <https://commits.webkit.org/r281913>

All reviewed patches have been landed. Closing bug and clearing flags on attachment 437119 [details].